1<!DOCTYPE html> 2<html> 3<head> 4 <meta charset="utf-8"> 5 <!-- Meta tags for social media banners, these should be filled in appropriatly as they are your "business card" --> 6 <!-- Replace the content tag with appropriate information --> 7 <meta name="description" content="DESCRIPTION META TAG"> 8 <meta property="og:title" content="SOCIAL MEDIA TITLE TAG"/> 9 <meta property="og:description" content="SOCIAL MEDIA DESCRIPTION TAG TAG"/> 10 <meta property="og:url" content="URL OF THE WEBSITE"/> 11 <!-- Path to banner image, should be in the path listed below. Optimal dimenssions are 1200X630--> 12 <meta property="og:image" content="static/image/your_banner_image.png" /> 13 <meta property="og:image:width" content="1200"/> 14 <meta property="og:image:height" content="630"/> 15 16 17 <meta name="twitter:title" content="TWITTER BANNER TITLE META TAG"> 18 <meta name="twitter:description" content="TWITTER BANNER DESCRIPTION META TAG"> 19 <!-- Path to banner image, should be in the path listed below. Optimal dimenssions are 1200X600--> 20 <meta name="twitter:image" content="static/images/your_twitter_banner_image.png"> 21 <meta name="twitter:card" content="summary_large_image"> 22 <!-- Keywords for your paper to be indexed by--> 23 <meta name="keywords" content="KEYWORDS SHOULD BE PLACED HERE"> 24 <meta name="viewport" content="width=device-width, initial-scale=1"> 25 26 27 <title>MaskUnet</title> 28 <link rel="icon" type="image/x-icon" href="static/images/icon-maskunet.png"> 29 <link href="https://fonts.googleapis.com/css?family=Google+Sans|Noto+Sans|Castoro" 30 rel="stylesheet"> 31 32 <link rel="stylesheet" href="static/css/bulma.min.css"> 33 <link rel="stylesheet" href="static/css/bulma-carousel.min.css"> 34 <link rel="stylesheet" href="static/css/bulma-slider.min.css"> 35 <link rel="stylesheet" href="static/css/fontawesome.all.min.css"> 36 <link rel="stylesheet" 37 href="https://cdn.jsdelivr.net/gh/jpswalsh/academicons@1/css/academicons.min.css"> 38 <link rel="stylesheet" href="static/css/index.css"> 39 40
40<script src="https://ajax.googleapis.com/ajax/libs/jquery/3.5.1/jquery.min.js"></script>
40 41
41<script src="https://documentcloud.adobe.com/view-sdk/main.js"></script>
41 42
42<script defer src="static/js/fontawesome.all.min.js"></script>
42 43
43<script src="static/js/bulma-carousel.min.js"></script>
43 44
44<script src="static/js/bulma-slider.min.js"></script>
44 45
45<script src="static/js/index.js"></script>
45 46</head> 47<body> 48 49 50 <section class="hero"> 51 <div class="hero-body"> 52 <div class="container is-max-desktop"> 53 <div class="columns is-centered"> 54 <div class="column has-text-centered"> 55 <h1 class="title is-1 publication-title"> 56 <img src="./static/images/icon-maskunet.png" alt="" style="height: 2.5em;"> 57 <br>Not All Parameters Matter: Masking Diffusion Models for Enhancing Generation Ability</h1> 58 <div class="is-size-5 publication-authors"> 59 <!-- Paper authors --> 60 <span class="author-block"> 61 <a href="https://gudaochangsheng.github.io/" target="_blank">Lei Wang</a><sup>1</sup> 62 </span> 63 64 <span class="author-block"> 65 <a href="https://sen-mao.github.io/" target="_blank">Senmao Li</a><sup>1</sup> 66 </span> 67 68 <span class="author-block"> 69 <a href="https://scholar.google.com/citations?hl=zh-CN&user=S1gksNwAAAAJ" target="_blank">Fei Yang</a><sup>1✉</sup> 70 </span> 71 72 <span class="author-block"> 73 <a href="" target="_blank">Jianye Wang</a><sup>1</sup> 74 </span> 75 76 <span class="author-block"> 77 <a href="" target="_blank">Ziheng Zhang</a><sup>1</sup> 78 </span> 79 80 <span class="author-block"> 81 <a href="" target="_blank">Yuhan Liu</a><sup>1</sup> 82 </span> 83 84 <span class="author-block"> 85 <a href="https://yaxingwang.github.io/" target="_blank">Yaxing Wang</a><sup>1,2</sup> 86 </span> 87 88 <span class="author-block"> 89 <a href="https://scholar.google.com.hk/citations?user=6CIDtZQAAAAJ&hl=en" target="_blank">Jian Yang</a><sup>1✉</sup> 90 </span> 91 </div> 92 93 <div class="is-size-5 publication-authors"> 94 <span class="author-block"><sup>1</sup>PCA Lab, VCIP, College of Computer Science, Nankai University, <sup>2</sup>Shenzhen Futian, NKIARI<br>CVPR 2025</span> 95 <span class="eql-cntrb"><small><br><sup>✉</sup>Corresponding authors</small></span> 96 </div> 97 98 <div class="column has-text-centered"> 99 <div class="publication-links"> 100 <!-- Arxiv PDF link --> 101 <span class="link-block"> 102 <a href="https://openaccess.thecvf.com/content/CVPR2025/papers/Wang_Not_All_Parameters_Matter_Masking_Diffusion_Models_for_Enhancing_Generation_CVPR_2025_paper.pdf" target="_blank" 103 class="external-link button is-normal is-rounded is-dark"> 104 <span class="icon"> 105 <i class="fas fa-file-pdf"></i> 106 </span> 107 <span>Paper</span> 108 </a> 109 </span> 110 111 <!-- Supplementary PDF link --> 112 <span class="link-block"> 113 <a href="static/pdfs/00985_supp.pdf" target="_blank" 114 class="external-link button is-normal is-rounded is-dark"> 115 <span class="icon"> 116 <i class="fas fa-file-pdf"></i> 117 </span> 118 <span>Supplementary</span> 119 </a> 120 </span> 121 122 <!-- Github link --> 123 <span class="link-block"> 124 <a href="https://github.com/gudaochangsheng/MaskUnet" target="_blank" 125 class="external-link button is-normal is-rounded is-dark"> 126 <span class="icon"> 127 <i class="fab fa-github"></i> 128 </span> 129 <span>Code</span> 130 </a> 131 </span> 132 133 <!-- ArXiv abstract Link -->
134 <span class="link-block"> 135 <a href="https://arxiv.org/abs/2505.03097" target="_blank" 136 class="external-link button is-normal is-rounded is-dark"> 137 <span class="icon"> 138 <i class="ai ai-arxiv"></i> 139 </span> 140 <span>arXiv</span> 141 </a> 142 </span> 143 </div> 144 </div> 145 </div> 146 </div> 147 </div> 148 </div> 149</section> 150 151 152<section class="hero is-small"> 153 <div class="container is-max-desktop"> 154 <div class="hero-body"> 155 <center> 156 <img src="static/motivation-mask.png"/> 157 </center> 158 <h2 class="subtitle has-text-centered" style="color: gray; font-size: 16px;"> 159 Analysis of parameter distributions and denoising effects across different time steps for Stable Diffusion (SD) 1.5 with and without random masking. 160The first column shows the parameter distribution of SD 1.5, while the second to fifth columns display the distributions of parameters removed by the 161random mask. The last two columns compare the generated samples from SD 1.5 and the random mask. 162 </h2> 163 </div> 164 </div> 165</section> 166 167<!-- Paper abstract --> 168<section class="section hero is-light"> 169 <div class="container is-max-desktop"> 170 <div class="columns is-centered has-text-centered"> 171 <div class="column is-four-fifths"> 172 <h2 class="title is-3">Abstract</h2> 173 <div class="content has-text-justified"> 174 <p> 175 The diffusion models, in early stages focus on constructing basic image structures, while the refined details, including local features and textures, are generated in later stages. Thus the same network layers are forced to learn both structural and textural information simultaneously, significantly differing from the traditional deep learning architectures (e.g., ResNet or GANs) which captures or generates the image semantic information at different layers. This difference inspires us to explore the time-wise diffusion models. We initially investigate the key contributions of the U-Net parameters to the denoising process and identify that properly zeroing out certain parameters (including large parameters) contributes to denoising, substantially improving the generation quality on the fly. Capitalizing on this discovery, we propose a simple yet effective methodâtermed âMaskUNetââ that enhances generation quality with negligible parameter numbers. 176Our method fully leverages timestep- and sample-dependent effective U-Net parameters. To optimize MaskUNet, we offer two fine-tuning strategies: a training-based approach and a training-free approach, including tailored networks and optimization functions. In zero-shot inference on the COCO dataset, MaskUNet achieves the best FID score and further demonstrates its effectiveness in downstream task evaluations. 177 </p> 178 </div> 179 </div> 180 </div> 181 </div> 182</section> 183<!-- End paper abstract --> 184 185 <section class="hero is-small"> 186 <div class="hero-body"> 187 <div class="container"> 188 <div class="columns is-centered has-text-centered"> 189 <h2 class="title is-3">Method<br></h2> 190 </div> 191 <br> 192 <div id="results-carousel"> 193 <div class="item"> 194 <!-- Your image here --> 195 <center> 196 <img src="static/images/method-maskunet.png" alt="MY ALT TEXT" style="width: 900px; height: auto;"/> 197 <figcaption style="width: 900px; height: auto;"> 198 <strong>The pipeline of the MaskUnet. G-Sig represents the Gumbel-Sigmoid activate function. GAP is global average pooling. 199 </figcaption> 200 </center> 201 </div> 202 </div> 203</div> 204</div> 205</section> 206 207 208<style> 209 .carousel .item img { 210 height: 800px; 211 width: auto; 212 display: block; 213 margin: 0 auto; 214 object-fit: contain; 215 } 216</style> 217 218<section class="hero is-small"> 219 <div class="hero-body"> 220 <div class="container"> 221 <div class="columns is-centered has-text-centered"> 222 <h2 class="title is-3">Results<br></h2> 223 </div> 224 <br> 225 <div id="results-carousel" class="carousel results-carousel"> 226 <div class="item"> 227 <img src="static/images/zero-shot-maskunet.png" alt="MY ALT TEXT" /> 228 <h2 class="subtitle has-text-centered"> 229 Quality results compared to other methods. 230 </h2> 231 </div> 232 <div class="item"> 233 <img src="static/images/Dreambooth.png" alt="MY ALT TEXT" /> 234 <h2 class="subtitle has-text-centered"> 235 Quality results compared to other methods. 236 </h2> 237 </div> 238 <div class="item"> 239 <img src="static/images/textual-inversion.png" alt="MY ALT TEXT" /> 240 <h2 class="subtitle has-text-centered"> 241 Quality results by Textual Inversion with or without mask. 242 </h2> 243 </div> 244 <div class="item"> 245 <img src="static/images/reversion.png" alt="MY ALT TEXT" /> 246 <h2 class="subtitle has-text-centered"> 247 Quality results by ReVersion with or without mask. 248 </h2> 249 </div> 250 <div class="item"> 251 <img src="static/images/t2v.png" alt="MY ALT TEXT" /> 252 <h2 class="subtitle has-text-centered"> 253 Quality results by Text2Video-Zero with or without mask. 254 </h2> 255 </div> 256 </div> 257 </div> 258 </div> 259</section> 260 261 262 263 264<section class="hero is-small"> 265 <div class="hero-body"> 266 <div class="container"> 267 <div class="columns is-centered has-text-centered"> 268 <h2 class="title is-3">Mask Analysis<br></h2> 269 </div> 270 <br> 271 <div id="results-carousel"> 272 <div class="item"> 273 <!-- Your image here --> 274 <center> 275 <img src="static/images/mask-ana.png" alt="MY ALT TEXT" style="width: 900px; height: auto;"/>
276 <figcaption style="width: 900px; height: auto;"> 277 <strong>(a) Visualization of image distributions for different methods using t-SNE. (b) Relationship between mask ratio and FID across checkpoint iterations. 278 </figcaption> 279 </center> 280 </div> 281 </div> 282</div> 283</div> 284</section> 285 286 287 288 289 290 291<!-- Paper poster --> 292<section class="hero is-small is-light"> 293 <div class="hero-body"> 294 <div class="container"> 295 <h2 class="title">Poster</h2> 296 297 <iframe src="static/pdfs/cvpr-poster.pdf" width="100%" height="550"> 298 </iframe> 299 300 </div> 301 </div> 302 </section> 303<!--End paper poster --> 304 305 306<!--BibTex citation --> 307 <section class="section" id="BibTeX"> 308 <div class="container is-max-desktop content"> 309 <h2 class="title">BibTeX</h2> 310 <pre><code>@inproceedings{wang2025not, 311 title={Not All Parameters Matter: Masking Diffusion Models for Enhancing Generation Ability}, 312 author={Wang, Lei and Li, Senmao and Yang, Fei and Wang, Jianye and Zhang, Ziheng and Liu, Yuhan and Wang, Yaxing and Yang, Jian}, 313 booktitle={Proceedings of the Computer Vision and Pattern Recognition Conference}, 314 pages={12880--12890}, 315 year={2025} 316}</code></pre> 317 </div> 318</section> 319<!--End BibTex citation --> 320 321 322 <footer class="footer"> 323 <div class="container"> 324 <div class="columns is-centered"> 325 <div class="column is-8"> 326 <div class="content"> 327 328 <p> 329 This page was built using the <a href="https://github.com/eliahuhorwitz/Academic-project-page-template" target="_blank">Academic Project Page Template</a> which was adopted from the <a href="https://nerfies.github.io" target="_blank">Nerfies</a> project page. 330 You are free to borrow the source code of this website, we just ask that you link back to this page in the footer. <br> This website is licensed under a <a rel="license" href="http://creativecommons.org/licenses/by-sa/4.0/" target="_blank">Creative 331 Commons Attribution-ShareAlike 4.0 International License</a>. 332 </p> 333 334 </div> 335 </div> 336 </div> 337 </div> 338</footer> 339 340<!-- Statcounter tracking code --> 341 342<!-- You can add a tracker to track page visits by creating an account at statcounter.com --> 343 344 <!-- End of Statcounter Code --> 345 346 </body> 347 </html>
Line numbers count LF bytes from the start of the resource, as the search results do. Vendor segments are library code the classifier recognised; they are stored but not indexed. Bytes are shown as Latin1 characters, one per byte.